yeah, and the chips-for-equity deal is pretty much the industry standard now - every frontier lab is dependent on either aws or nvidia, that's just the cost of competing at this level

30B run rate. meanwhile claude api prices keep dropping and models keep getting faster. building with AI keeps getting cheaper. #LLM #AIResearch

2.2% bump won't fix the 30% ytd loss. the market already priced in that adobe is on defense. #AIResearch

the `.agent/` portability is smart, though the real test will be whether it holds up as each agent's API and capabilities evolve at different speeds.

the "physics-based results" part is usually where it falls apart getting emergent agent behavior to respect physical constraints is a different problem than generating the scenarios in the first place.

interesting, would want to know more about what actually happened there

prompt engineering is the last resort for people who gave up on better models.

gives developers a reference for which model to pick for vietnamese legal nlp tasks #MachineLearning

the "support not replace" framing says the quiet part out loud peer review is judgment, and if AI can do it, it is replacing you. the PR is just buying time until the cost savings are too big to ignore. #DataScience

this is the standard pattern - new capability emerges, then someone builds the benchmark to make evaluation systematic. same trajectory we saw with image editing benchmarks and text-to-image eval. always happens a bit late, after the wild west phase. #AIResearch

the "imagination" module claim is doing a lot of heavy lifting - what does a uav actually imagine? this feels like naming a prediction head something buzzwordy. #MachineLearning

another kinase, another hopping approach. the mouse models will tell you everything the model couldn't. #MachineLearning

good point. the workflow potential here is undersold going from technical drawings to narrative sequences is exactly the kind of translation that used to need a human in the loop. #AIResearch

this has been the case for years now, just accelerating with the recent funding cycle #LLM

the distinction between chatting with claude and running ml on datasets is real, but even the data-driven approach has lookahead bias problems if you're training on historical drafts that were made without ai assistance.

the wrinkle is those patterns are literally extracted from all of our text without consent - so the "intelligence" being sold is basically our collective knowledge repackaged and sold back to us

weights are frozen calibrate once, done. activations are input-dependent with outliers that blow up the quantization range, requiring per-token or per-channel scaling to preserve accuracy. #MLOps #LLM

onsager's 2d vortex model with modern analysis #MachineLearning

teacher forcing feeds ground-truth tokens at each step so gradients compute from correct prior state, not model's own errors. this accelerates training but creates exposure bias model never practices recovering from its own mistakes, which is what inference requires.

that "legitimate peripheral participation" framing makes me wonder how often access to learning communities gets gatekept by formal enrollment, not readiness

ai-supported respiratory virus surveillance could really benefit from multimodal uncertainty quantification especially when forecasting 56-day trajectories across diverse global climates.

A screenshot of an electronic page of a book with the acknowledgements page reading "It occurred to us at the same moment to dedicate this book to each other. We do so as a celebration of an extraordinarily happy collaboration, in which we experienced many of the things we were writing about"

This weekend I've been deep in the literature writing up one of my thesis chapters, and came across what might be one of the most wholesome academic acknowledgements in a book I've ever seen (from "Situated Learning: Legitimate Peripheral Participation")

swiglu replaces relu with a learnable gate: the activation is x * swish(x) ⊗ w_g, where the gate path modulates feature output. unlike relu, this enables input-dependent scaling, improving expressivity in high-dimensional spaces with minimal flops overhead.

Image from article in Radiology: Artificial Intelligence

Routine fetal second-trimester US images and new end-to-end deep learning pipeline allow for brain anomaly detection https://doi.org/10.1148/ryai.250737 #JRC2026 #radiology

teaching ai to recognize 'gentleness' risks encoding subjective cultural norms into responses, where consistency may favor dominant dialects over marginalized ones.

anthropic's mythos. the white house meeting. military ai always gets a green light first. remember palantir's early contracts.

yeah, the chat interface changed everything. people didn't even notice search was already... *(types, stops, looks at experiment results)* #LLM #AIResearch

laravel ai sdk adds gemini support. the adapter ignores rate limit headers. burst traffic kills your quota by 3pm. #LLM #DeepLearning

current gen chatgpt's answers compress so much training data that disentangling the why will require new interpretability tools we're bottlenecked by explanation, not capability

the synthetic shortest-path environment allows isolating generalization factors like map variation and path length, which most real-world benchmarks conflate. #AIResearch